Skip to content

Use Claude Code, Codex, VSCode, Pi, and OpenCode (and 6 other harnesses) for free (1.3B+ free tokens) from your terminal, app, IDE, or phone, and now from the browser with native browser sessions (multi-harness + multi-model) like OpenClaw (voice supported + ToS friendly)

Python

Preserve mid-conversation system messages through provider conversion (#1125)

Ali Khokharf0b31065eef5
diff --git a/ARCHITECTURE.md b/ARCHITECTURE.md
index 463f0157..109cc5f5 100644
--- a/ARCHITECTURE.md
+++ b/ARCHITECTURE.md
@@ -644,6 +644,14 @@ usage quirks such as DeepSeek prompt-cache counters.
stream ledger, Anthropic SSE emitter, continuation-body construction, and tool repair;
- token counting and Anthropic-owned failure-kind-to-wire mapping.
+Anthropic request models validate transcript data without merging, hoisting, or
+reordering semantically meaningful message roles. Top-level `system` content
+stays distinct from inline `system` messages. Target-protocol conversion owns
+their representation: neutral OpenAI Chat conversion preserves inline role and
+order, and rejects unrepresentable blocks instead of dropping them. Any
+provider-specific deviation belongs in an explicit request policy backed by a
+known upstream incompatibility.
+
User image conversion is a pure protocol operation. Core maps Anthropic base64
and URL image sources to ordered OpenAI `image_url` content parts without
fetching remote content. Provider adapters do not gate that conversion behind a
@@ -772,6 +780,10 @@ common low-value client requests before they reach a provider:
- suggestion mode;
- filepath extraction.
+Detection derives a read-only semantic view: inline `system` messages contribute
+system context but are not counted as conversational turns. The original
+request remains ordered and unchanged for provider execution.
+
The Messages handler runs these only after model routing and after local server-tool
handling. Each optimization is controlled by settings flags.
diff --git a/pyproject.toml b/pyproject.toml
index b838a384..07b2bd85 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
[project]
name = "free-claude-code"
-version = "4.6.1"
+version = "4.6.3"
description = "Local proxy connecting coding agents to OpenAI-compatible AI providers"
readme = "README.md"
requires-python = ">=3.14.0"
diff --git a/src/free_claude_code/api/detection.py b/src/free_claude_code/api/detection.py
index 1dfbc81f..1bac08e4 100644
--- a/src/free_claude_code/api/detection.py
+++ b/src/free_claude_code/api/detection.py
@@ -4,7 +4,39 @@ Detects quota checks, title generation, prefix detection, safety classifier,
suggestion mode, and filepath extraction requests to enable targeted handling.
"""
-from free_claude_code.core.anthropic import MessagesRequest, extract_text_from_content
+from free_claude_code.core.anthropic import (
+ Message,
+ MessagesRequest,
+ extract_text_from_content,
+)
+
+
+def _single_user_turn(request_data: MessagesRequest) -> Message | None:
+ """Return the sole conversational user turn, ignoring system context."""
+ user_turn: Message | None = None
+ for message in request_data.messages:
+ if message.role == "system":
+ continue
+ if message.role != "user" or user_turn is not None:
+ return None
+ user_turn = message
+ return user_turn
+
+
+def _request_system_text(request_data: MessagesRequest) -> str:
+ """Return top-level and inline system text for request-shape detection."""
+ parts: list[str] = []
+ if request_data.system:
+ text = extract_text_from_content(request_data.system)
+ if text:
+ parts.append(text)
+ for message in request_data.messages:
+ if message.role != "system":
+ continue
+ text = extract_text_from_content(message.content)
+ if text:
+ parts.append(text)
+ return "\n".join(parts)
def is_quota_check_request(request_data: MessagesRequest) -> bool:
@@ -13,15 +45,11 @@ def is_quota_check_request(request_data: MessagesRequest) -> bool:
Quota checks are typically simple requests with max_tokens=1
and a single message containing the word "quota".
"""
- if (
- request_data.max_tokens == 1
- and len(request_data.messages) == 1
- and request_data.messages[0].role == "user"
- ):
- text = extract_text_from_content(request_data.messages[0].content)
- if "quota" in text.lower():
- return True
- return False
+ message = _single_user_turn(request_data)
+ if request_data.max_tokens != 1 or message is None:
+ return False
+ text = extract_text_from_content(message.content)
+ return "quota" in text.lower()
def is_title_generation_request(request_data: MessagesRequest) -> bool:
@@ -33,9 +61,9 @@ def is_title_generation_request(request_data: MessagesRequest) -> bool:
Matches Claude Code session title prompts (sentence-case title, JSON
\"title\" field, etc.).
"""
- if not request_data.system or request_data.tools:
+ if request_data.tools:
return False
- system_text = extract_text_from_content(request_data.system).lower()
+ system_text = _request_system_text(request_data).lower()
if "title" not in system_text:
return False
return "sentence-case title" in system_text or (
@@ -54,10 +82,11 @@ def is_prefix_detection_request(request_data: MessagesRequest) -> tuple[bool, st
Returns:
Tuple of (is_prefix_request, command_string)
"""
- if len(request_data.messages) != 1 or request_data.messages[0].role != "user":
+ message = _single_user_turn(request_data)
+ if message is None:
return False, ""
- content = extract_text_from_content(request_data.messages[0].content)
+ content = extract_text_from_content(message.content)
if "<policy_spec>" in content and "Command:" in content:
try:
@@ -111,12 +140,13 @@ def is_filepath_extraction_request(
Returns:
Tuple of (is_filepath_request, command, output)
"""
- if len(request_data.messages) != 1 or request_data.messages[0].role != "user":
+ message = _single_user_turn(request_data)
+ if message is None:
return False, "", ""
if request_data.tools:
return False, "", ""
- content = extract_text_from_content(request_data.messages[0].content)
+ content = extract_text_from_content(message.content)
if "Command:" not in content or "Output:" not in content:
return False, "", ""
@@ -125,9 +155,7 @@ def is_filepath_extraction_request(
user_has_filepaths = (
"filepaths" in content.lower() or "<filepaths>" in content.lower()
)
- system_text = (
- extract_text_from_content(request_data.system) if request_data.system else ""
- )
+ system_text = _request_system_text(request_data)
system_has_extract = (
"extract any file paths" in system_text.lower()
or "file paths that this command" in system_text.lower()
diff --git a/src/free_claude_code/core/anthropic/__init__.py b/src/free_claude_code/core/anthropic/__init__.py
index 699254a2..cf7a2a09 100644
--- a/src/free_claude_code/core/anthropic/__init__.py
+++ b/src/free_claude_code/core/anthropic/__init__.py
@@ -27,7 +27,6 @@ from .models import (
Message,
MessagesRequest,
MessagesResponse,
- Role,
SystemContent,
ThinkingConfig,
TokenCountRequest,
@@ -71,7 +70,6 @@ __all__ = [
"MessagesResponse",
"OpenAIConversionError",
"ReasoningReplayMode",
- "Role",
"StreamBlockLedger",
"SystemContent",
"ThinkTagParser",
diff --git a/src/free_claude_code/core/anthropic/conversion.py b/src/free_claude_code/core/anthropic/conversion.py
index ef4f7ec8..eb07ac60 100644
--- a/src/free_claude_code/core/anthropic/conversion.py
+++ b/src/free_claude_code/core/anthropic/conversion.py
@@ -164,6 +164,34 @@ def _assert_no_forbidden_assistant_block(block: Any) -> None:
)
+def _openai_system_text(
+ content: Any,
+ *,
+ context: str,
+) -> str | None:
+ """Return OpenAI-compatible system text without silently dropping blocks."""
+ if isinstance(content, str):
+ return content
+ if not isinstance(content, list):
+ raise OpenAIConversionError(
+ f"OpenAI chat conversion requires {context} content to be text."
+ )
+ if not content:
+ return None
+
+ text_parts: list[str] = []
+ for block in content:
+ block_type = get_block_type(block)
+ if block_type != "text":
+ raise OpenAIConversionError(
+ f"OpenAI chat conversion cannot represent {context} content block "
+ f"{block_type!r} without data loss."
+ )
+ text_parts.append(str(get_block_attr(block, "text", "")))
+
+ return "\n\n".join(text_parts)
+
+
def _openai_user_image_part(block: Any) -> dict[str, Any]:
"""Convert one Anthropic user image block without performing I/O."""
source = get_block_attr(block, "source", {})
@@ -364,6 +392,17 @@ class AnthropicToOpenAIConverter:
reasoning_content: str | None,
reasoning_replay: ReasoningReplayMode,
) -> list[_TranscriptSegment]:
+ if role == "system":
+ system_text = _openai_system_text(
+ content,
+ context="an inline Anthropic system message",
+ )
+ if system_text is None:
+ raise OpenAIConversionError(
+ "OpenAI chat conversion requires an inline Anthropic system "
+ "message to contain text."
+ )
+ return [_PlainSegment([{"role": "system", "content": system_text}])]
if role == "assistant" and isinstance(content, list):
if (first_i := _index_first_tool_use(content)) is not None:
for block in content:
@@ -602,17 +641,17 @@ class AnthropicToOpenAIConverter:
@staticmethod
def convert_system_prompt(system: Any) -> dict[str, str] | None:
- if isinstance(system, str):
- return {"role": "system", "content": system}
+ if system is None:
+ return None
+ system_text = _openai_system_text(
+ system,
+ context="the top-level Anthropic system prompt",
+ )
+ if system_text is None:
+ return None
if isinstance(system, list):
- text_parts = [
- get_block_attr(block, "text", "")
- for block in system
- if get_block_type(block) == "text"
- ]
- if text_parts:
- return {"role": "system", "content": "\n\n".join(text_parts).strip()}
- return None
+ system_text = system_text.strip()
+ return {"role": "system", "content": system_text}
def build_base_request_body(
diff --git a/src/free_claude_code/core/anthropic/models.py b/src/free_claude_code/core/anthropic/models.py
index a0f639da..684f8cb0 100644
--- a/src/free_claude_code/core/anthropic/models.py
+++ b/src/free_claude_code/core/anthropic/models.py
@@ -1,15 +1,8 @@
"""Pydantic models for the Anthropic Messages protocol."""
-from enum import StrEnum
from typing import Any, Literal
-from pydantic import BaseModel, ConfigDict, Field, model_validator
-
-
-class Role(StrEnum):
- user = "user"
- assistant = "assistant"
- system = "system"
+from pydantic import BaseModel, ConfigDict, Field
class _AnthropicBlockBase(BaseModel):
@@ -85,58 +78,8 @@ class SystemContent(_AnthropicBlockBase):
text: str
-def _text_system_block(text: str) -> dict[str, str]:
- return {"type": "text", "text": text}
-
-
-def _merge_system_values(existing: Any, additions: list[Any]) -> Any:
- values = [existing] if existing is not None else []
- values.extend(additions)
-
- if all(isinstance(value, str) for value in values):
- return "\n\n".join(value for value in values if value)
-
- blocks: list[Any] = []
- for value in values:
- if isinstance(value, str):
- blocks.append(_text_system_block(value))
- elif isinstance(value, list):
- blocks.extend(value)
- else:
- blocks.append(value)
- return blocks
-
-
-def _normalize_system_role_messages(data: Any) -> Any:
- if not isinstance(data, dict):
- return data
-
- messages = data.get("messages")
- if not isinstance(messages, list):
- return data
-
- system_contents: list[Any] = []
- normalized_messages: list[Any] = []
- for message in messages:
- role = message.get("role") if isinstance(message, dict) else None
- if role == Role.system:
- system_contents.append(message.get("content", ""))
- continue
- normalized_messages.append(message)
-
- if not system_contents:
- return data
-
- normalized = dict(data)
- normalized["messages"] = normalized_messages
- normalized["system"] = _merge_system_values(
- normalized.get("system"), system_contents
- )
- return normalized
-
-
class Message(BaseModel):
- role: Literal["user", "assistant"]
+ role: Literal["user", "assistant", "system"]
content: (
str
| list[
@@ -171,11 +114,6 @@ class ThinkingConfig(BaseModel):
class MessagesRequest(BaseModel):
model_config = ConfigDict(extra="allow")
- @model_validator(mode="before")
- @classmethod
- def normalize_system_role_messages(cls, data: Any) -> Any:
- return _normalize_system_role_messages(data)
-
model: str
original_model: str | None = Field(default=None, exclude=True)
resolved_provider_model: str | None = Field(default=None, exclude=True)
@@ -201,11 +139,6 @@ class MessagesRequest(BaseModel):
class TokenCountRequest(BaseModel):
model_config = ConfigDict(extra="allow")
- @model_validator(mode="before")
- @classmethod
- def normalize_system_role_messages(cls, data: Any) -> Any:
- return _normalize_system_role_messages(data)
-
model: str
original_model: str | None = Field(default=None, exclude=True)
resolved_provider_model: str | None = Field(default=None, exclude=True)
diff --git a/tests/api/test_api.py b/tests/api/test_api.py
index 6902d84a..e8d51680 100644
--- a/tests/api/test_api.py
+++ b/tests/api/test_api.py
@@ -235,8 +235,8 @@ def test_create_message_pre_start_provider_error_returns_terminal_json(
mock_provider.stream_response = _mock_stream_response
-def test_create_message_accepts_system_role_messages(client: TestClient):
- """Create message accepts latest-client system messages."""
+def test_create_message_preserves_system_role_messages(client: TestClient):
+ """Create message preserves latest-client system message placement."""
mock_provider.stream_response = _mock_stream_response
_stream_response_calls.clear()
payload = {
@@ -254,8 +254,13 @@ def test_create_message_accepts_system_role_messages(client: TestClient):
assert response.status_code == 200
routed_request = _stream_response_calls[0][0][0]
- assert [message.role for message in routed_request.messages] == ["user", "user"]
- assert routed_request.system == "system prompt"
+ assert [message.role for message in routed_request.messages] == [
+ "user",
+ "system",
+ "user",
+ ]
+ assert routed_request.messages[1].content == "system prompt"
+ assert routed_request.system is None
def test_model_mapping(client: TestClient):
diff --git a/tests/api/test_detection.py b/tests/api/test_detection.py
index 800e65b5..3fbfaff0 100644
--- a/tests/api/test_detection.py
+++ b/tests/api/test_detection.py
@@ -5,21 +5,57 @@ from unittest.mock import patch
from free_claude_code.api.detection import (
is_filepath_extraction_request,
is_prefix_detection_request,
+ is_quota_check_request,
is_safety_classifier_request,
+ is_title_generation_request,
)
from free_claude_code.core.anthropic.models import Message, MessagesRequest
-def _make_request(content: str, **kwargs) -> MessagesRequest:
+def _make_request(
+ content: str, *, inline_system: str | None = None, **kwargs
+) -> MessagesRequest:
+ messages = []
+ if inline_system is not None:
+ messages.append(Message(role="system", content=inline_system))
+ messages.append(Message(role="user", content=content))
return MessagesRequest(
model="claude-3-sonnet",
- max_tokens=100,
- messages=[Message(role="user", content=content)],
+ max_tokens=kwargs.pop("max_tokens", 100),
+ messages=messages,
**kwargs,
)
+def test_quota_detection_ignores_inline_system_context() -> None:
+ request = _make_request(
+ "Check my quota", inline_system="Current request context", max_tokens=1
+ )
+
+ assert is_quota_check_request(request) is True
+
+
+def test_title_detection_reads_inline_system_context() -> None:
+ request = _make_request(
+ "Summarize this session",
+ inline_system=(
+ "Generate a concise, sentence-case title for this coding session. "
+ 'Return JSON with a single "title" field.'
+ ),
+ )
+
+ assert is_title_generation_request(request) is True
+
+
class TestIsPrefixDetectionRequest:
+ def test_inline_system_context_does_not_hide_single_user_turn(self):
+ req = _make_request(
+ "<policy_spec> Command: git status",
+ inline_system="Current request context",
+ )
+
+ assert is_prefix_detection_request(req) == (True, "git status")
+
def test_output_marker_handling(self):
"""Content with Command: but Output: after cmd_start; output has < or \\n\\n."""
content = "<policy_spec> Command:\nls -la\nOutput:\na.txt\nb.txt\n\nmore"
@@ -95,6 +131,14 @@ class TestIsSafetyClassifierRequest:
class TestIsFilepathExtractionRequest:
+ def test_inline_system_context_preserves_filepath_detection(self):
+ req = _make_request(
+ "Command:\nls\nOutput:\na.txt",
+ inline_system="Extract any file paths that this command reads or modifies.",
+ )
+
+ assert is_filepath_extraction_request(req) == (True, "ls", "a.txt")
+
def test_output_marker_minus_one_returns_false(self):
"""Output: not found after Command: returns False."""
content = "Command:\nls\nfilepaths"
diff --git a/tests/api/test_request_utils.py b/tests/api/test_request_utils.py
index 6a1d3101..d0126094 100644
--- a/tests/api/test_request_utils.py
+++ b/tests/api/test_request_utils.py
@@ -130,6 +130,7 @@ class TestTitleGenerationRequest:
req = MagicMock(spec=MessagesRequest)
req.system = self._title_gen_system()
req.tools = None
+ req.messages = []
assert is_title_generation_request(req) is True
@@ -138,6 +139,7 @@ class TestTitleGenerationRequest:
req = MagicMock(spec=MessagesRequest)
req.system = self._title_gen_system()
req.tools = [MagicMock()]
+ req.messages = []
assert is_title_generation_request(req) is False
@@ -146,6 +148,7 @@ class TestTitleGenerationRequest:
req = MagicMock(spec=MessagesRequest)
req.system = None
req.tools = None
+ req.messages = []
assert is_title_generation_request(req) is False
@@ -156,6 +159,7 @@ class TestTitleGenerationRequest:
req = MagicMock(spec=MessagesRequest)
req.system = [block]
req.tools = None
+ req.messages = []
assert is_title_generation_request(req) is False
@@ -166,6 +170,7 @@ class TestTitleGenerationRequest:
req = MagicMock(spec=MessagesRequest)
req.system = [block]
req.tools = None
+ req.messages = []
assert is_title_generation_request(req) is True
@@ -454,6 +459,16 @@ class TestGetTokenCount:
# Double message should have more tokens (including overhead)
assert count_double > count_single
+ def test_inline_system_message_contributes_to_token_count(self):
+ """Mid-conversation system content remains part of the counted transcript."""
+ user_message = Message(role="user", content="Hello")
+ inline_system = Message(role="system", content="New instructions")
+
+ without_inline_system = get_token_count([user_message])
+ with_inline_system = get_token_count([user_message, inline_system])
+
+ assert with_inline_system > without_inline_system
+
def test_per_message_overhead_four_tokens(self):
"""Per-message overhead is 4 tokens (was 3)."""
msg = MagicMock()
diff --git a/tests/core/anthropic/test_models.py b/tests/core/anthropic/test_models.py
index 01e26751..39f789d9 100644
--- a/tests/core/anthropic/test_models.py
+++ b/tests/core/anthropic/test_models.py
@@ -37,7 +37,7 @@ def test_messages_request_rejects_null_stream() -> None:
)
-def test_messages_request_normalizes_system_role_messages():
+def test_messages_request_preserves_system_role_message_order():
request = MessagesRequest.model_validate(
{
"model": "claude-3-opus",
@@ -50,11 +50,16 @@ def test_messages_request_normalizes_system_role_messages():
}
)
- assert [message.role for message in request.messages] == ["user", "user"]
- assert request.system == "system prompt"
+ assert [message.role for message in request.messages] == [
+ "user",
+ "system",
+ "user",
+ ]
+ assert request.messages[1].content == "system prompt"
+ assert request.system is None
-def test_messages_request_merges_system_role_messages_with_existing_system():
+def test_messages_request_keeps_top_level_and_inline_system_content_distinct():
request = MessagesRequest.model_validate(
{
"model": "claude-3-opus",
@@ -67,11 +72,12 @@ def test_messages_request_merges_system_role_messages_with_existing_system():
}
)
- assert len(request.messages) == 1
- assert request.system == "existing system\n\nmessage system"
+ assert request.system == "existing system"
+ assert [message.role for message in request.messages] == ["system", "user"]
+ assert request.messages[0].content == "message system"
-def test_messages_request_preserves_system_block_cache_control_when_normalizing():
+def test_messages_request_preserves_inline_system_block_metadata():
request = MessagesRequest.model_validate(
{
"model": "claude-3-opus",
@@ -99,14 +105,17 @@ def test_messages_request_preserves_system_block_cache_control_when_normalizing(
}
)
- assert len(request.messages) == 1
+ assert len(request.messages) == 2
assert isinstance(request.system, list)
- assert [block.text for block in request.system] == [
- "existing system",
- "message system",
- ]
+ assert [block.text for block in request.system] == ["existing system"]
assert request.system[0].model_dump()["cache_control"] == {"type": "ephemeral"}
- assert request.system[1].model_dump()["cache_control"] == {"type": "ephemeral"}
+ inline_content = request.messages[0].content
+ assert isinstance(inline_content, list)
+ assert inline_content[0].model_dump() == {
+ "type": "text",
+ "text": "message system",
+ "cache_control": {"type": "ephemeral"},
+ }
def test_messages_request_ignores_internal_routing_fields_when_supplied():
@@ -133,7 +142,7 @@ def test_token_count_request_parses_without_model_mapping_side_effects():
assert request.model == "claude-3-sonnet"
-def test_token_count_request_normalizes_system_role_messages():
+def test_token_count_request_preserves_system_role_messages():
request = TokenCountRequest.model_validate(
{
"model": "claude-3-sonnet",
@@ -144,9 +153,9 @@ def test_token_count_request_normalizes_system_role_messages():
}
)
- assert len(request.messages) == 1
- assert request.messages[0].role == "user"
- assert request.system == "counting system"
+ assert [message.role for message in request.messages] == ["system", "user"]
+ assert request.messages[0].content == "counting system"
+ assert request.system is None
def test_messages_request_preserves_thinking_signature():
diff --git a/tests/providers/test_converter.py b/tests/providers/test_converter.py
index 4e2ed077..23124bc9 100644
--- a/tests/providers/test_converter.py
+++ b/tests/providers/test_converter.py
@@ -59,6 +59,171 @@ def test_convert_system_prompt_none():
assert AnthropicToOpenAIConverter.convert_system_prompt(None) is None
+def test_openai_build_preserves_top_level_and_inline_system_message_order() -> None:
+ request = MessagesRequest.model_validate(
+ {
+ "model": "model",
+ "system": "Conversation-wide instructions",
+ "messages": [
+ {"role": "user", "content": "First question"},
+ {"role": "assistant", "content": "First answer"},
+ {"role": "system", "content": "Instructions from this point"},
+ {"role": "user", "content": "Second question"},
+ ],
+ }
+ )
+
+ body = build_base_request_body(request)
+
+ assert body["messages"] == [
+ {"role": "system", "content": "Conversation-wide instructions"},
+ {"role": "user", "content": "First question"},
+ {"role": "assistant", "content": "First answer"},
+ {"role": "system", "content": "Instructions from this point"},
+ {"role": "user", "content": "Second question"},
+ ]
+
+
+def test_openai_build_joins_inline_system_text_blocks_without_repositioning() -> None:
+ request = MessagesRequest.model_validate(
+ {
+ "model": "model",
+ "messages": [
+ {"role": "user", "content": "Before"},
+ {
+ "role": "system",
+ "content": [
+ {"type": "text", "text": "First instruction"},
+ {"type": "text", "text": "Second instruction"},
+ ],
+ },
+ {"role": "user", "content": "After"},
+ ],
+ }
+ )
+
+ body = build_base_request_body(request)
+
+ assert body["messages"][1] == {
+ "role": "system",
+ "content": "First instruction\n\nSecond instruction",
+ }
+
+
+def test_inline_system_message_preserves_existing_openai_cache_prefix() -> None:
+ prefix_request = MessagesRequest.model_validate(
+ {
+ "model": "model",
+ "system": "Conversation-wide instructions",
+ "messages": [
+ {"role": "user", "content": "First question"},
+ {"role": "assistant", "content": "First answer"},
+ ],
+ }
+ )
+ continued_request = MessagesRequest.model_validate(
+ {
+ "model": "model",
+ "system": "Conversation-wide instructions",
+ "messages": [
+ {"role": "user", "content": "First question"},
+ {"role": "assistant", "content": "First answer"},
+ {"role": "system", "content": "Instructions from this point"},
+ {"role": "user", "content": "Second question"},
+ ],
+ }
+ )
+
+ prefix = build_base_request_body(prefix_request)["messages"]
+ continued = build_base_request_body(continued_request)["messages"]
+
+ assert continued[: len(prefix)] == prefix
+ assert continued[len(prefix) :] == [
+ {"role": "system", "content": "Instructions from this point"},
+ {"role": "user", "content": "Second question"},
+ ]
+
+
+def test_inline_system_message_follows_completed_tool_result() -> None:
+ request = MessagesRequest.model_validate(
+ {
+ "model": "model",
+ "messages": [
+ {
+ "role": "assistant",
+ "content": [
+ {
+ "type": "tool_use",
+ "id": "call_1",
+ "name": "Read",
+ "input": {},
+ }
+ ],
+ },
+ {
+ "role": "user",
+ "content": [
+ {
+ "type": "tool_result",
+ "tool_use_id": "call_1",
+ "content": "done",
+ }
+ ],
+ },
+ {"role": "system", "content": "New instructions"},
+ ],
+ }
+ )
+
+ body = build_base_request_body(request)
+
+ assert [message["role"] for message in body["messages"]] == [
+ "assistant",
+ "tool",
+ "system",
+ ]
+
+
+def test_openai_build_rejects_non_text_inline_system_blocks() -> None:
+ request = MessagesRequest.model_validate(
+ {
+ "model": "model",
+ "messages": [
+ {
+ "role": "system",
+ "content": [
+ {
+ "type": "image",
+ "source": {
+ "type": "url",
+ "url": "https://example.com/a.png",
+ },
+ }
+ ],
+ }
+ ],
+ }
+ )
+
+ with pytest.raises(
+ OpenAIConversionError,
+ match="inline Anthropic system message content block 'image' without data loss",
+ ):
+ build_base_request_body(request)
+
+
+def test_openai_build_rejects_empty_inline_system_content() -> None:
+ request = MessagesRequest.model_validate(
+ {
+ "model": "model",
+ "messages": [{"role": "system", "content": []}],
+ }
+ )
+
+ with pytest.raises(OpenAIConversionError, match="contain text"):
+ build_base_request_body(request)
+
+
# --- Tool Conversion Tests ---
diff --git a/uv.lock b/uv.lock
index fcd4d290..57b3e703 100644
--- a/uv.lock
+++ b/uv.lock
@@ -561,7 +561,7 @@ wheels = [
[[package]]
name = "free-claude-code"
-version = "4.6.1"
+version = "4.6.3"
source = { editable = "." }
dependencies = [
{ name = "aiohttp" },